Forkcast Hacker News #11: turbo-fieldfare, cycloidal_gearbox, Swiftlet, kakehashi, gander, shitty...

Episode Summary#

Episode 11 covers 30 open-source projects from Hacker News, led by turbo-fieldfare (4990 stars) for running Gemma 4 26B in 2GB RAM on Mac. Also featured are cycloidal_gearbox (82 stars) built by a 15-year-old, Swiftlet (374 stars) running 80B models on 4.3GB RAM, kakehashi (350 stars) for macOS-on-Linux-ARM translation, and 26 more projects spanning developer tools, AI, security, and specialty projects.

Repositories Covered#

turbo-fieldfare#

Running large language models like Gemma 4 26B locally requires massive RAM, typically 16GB or more, making it impossible on most consumer hardware without expensive GPUs. Open-source engine running Gemma 4 26B in just 2 GB RAM on any M-series Mac. Uses fieldfare compression technique. Swift-native implementation. Optimized for Apple Silicon memory architecture. Achieves real-time inference on consumer laptops.

cycloidal_gearbox#

Cycloidal gearbox design and simulation tools are either expensive commercial software or require deep mechanical engineering expertise to use from scratch. Built by a 15-year-old engineer. Python-based cycloidal gearbox design and simulation. Includes 3D modeling, kinematic analysis, and manufacturing-ready output. CAD-compatible exports. Open-source alternative to commercial gear design tools.

Swiftlet#

Running large language models like 80B parameter Qwen requires massive hardware. Even on Mac, these models typically need 32GB+ RAM and run slowly. Run an 80B Qwen model in just 4.3 GB of RAM on a Mac. Also runs a 35B model on iPhone. Swift-native implementation using advanced quantization. Real-time inference on consumer devices. Memory-mapped model loading.

kakehashi#

Running macOS binaries on Linux ARM (like Apple Silicon cloud instances or Raspberry Pi) requires complex translation layers that are slow and incompatible with many applications. Experimental userspace to run macOS binaries on Linux ARM. Rust-based binary translation. Supports Mach-O to ELF translation. Handles syscalls and framework calls. Performance within 2-3x of native for many workloads.

gander#

Android file viewers either request excessive permissions, show ads, or lack basic features like proper archive and document handling. Android file viewer that asks for zero permissions. Supports archives, documents, images, videos, and code files. Built with React Native. Material Design 3 UI. No ads, no tracking, no network access. Open-source alternative to bloated file managers.

shitty#

Terminal emulators are either feature-rich but slow, or fast but lack modern conveniences like tabs, splits, and proper Unicode support. Deliberately memory-unsafe terminal emulator written for speed. Faster than most existing terminals. Python-based with C extensions for performance-critical paths. Supports tabs, splits, Unicode, and 256 colors. Minimal resource usage.

nixos-dgx-spark#

NVIDIA’s DGX Spark comes with Ubuntu but developers who prefer NixOS have no easy way to run it on this AI-focused hardware without losing key features. Nix and NixOS configuration for NVIDIA DGX Spark. Reproducible AI development environment. CUDA, cuDNN, and ML frameworks pre-configured. Flake-based for reproducibility. Community-maintained hardware support.

Soup#

Fine-tuning even small LLMs like 8B models typically requires 16GB+ VRAM GPUs, making it inaccessible to developers with consumer hardware. Fine-tune an 8B parameter model on a 4 GB laptop GPU. Uses aggressive memory optimization techniques. Gradient checkpointing, 4-bit quantization, and CPU offloading. Supports LoRA and QLoRA. Train on consumer laptops with integrated or low-end GPUs.

nightcrawler#

Penetration testing tools are complex, expensive, and require powerful hardware. Running a full pentest toolkit on a smartphone seems impossible. Local AI pentesting agent running on a smartphone. Uses LLM for automated vulnerability scanning and exploitation. Supports network scanning, web app testing, and privilege escalation. Runs entirely offline on mobile devices. Python-based with Android integration.

creakwork12#

Laptop speaker drivers and firmware are closed-source, making it impossible to customize or improve audio output through software. Makes Framework 12 laptop sound like a creaky door. Rust-based audio driver modification. Custom waveform synthesis. Real-time audio processing. Fun hardware experiment showing open firmware possibilities.

syncular#

SQL database sync between devices is complex, requiring heavy infrastructure like CouchDB or custom CRDT implementations that add overhead. Offline-first SQL sync with TypeScript and Rust cores. CRDT-based conflict resolution. Real-time sync between devices. Works with SQLite and PostgreSQL. Sub-millisecond merge times. No server required for P2P sync.

fec#

Viterbi decoders for error correction are critical in communications but existing implementations are either slow or not SIMD-optimized for modern hardware. SIMD-optimized Viterbi decoder in Rust. Uses AVX2 and NEON intrinsics for parallel decoding. Processes multiple trellis paths simultaneously. Suitable for software-defined radio and communications. Benchmark-proven speedup over scalar implementations.

katharos#

Python’s async model with asyncio is complex and error-prone. CSP-style concurrency (like Go channels) would be simpler for many use cases but isn’t natively available. Functional programming and CSP-style concurrency for Python. Go-like channels and goroutines. Pure functional data structures. Actor model support. Type-safe message passing. Built on top of asyncio with a cleaner API.

qapla#

Small language models (SLMs) for edge devices require expensive hardware. Running AI on a microcontroller costing just $8 seems impossible with current tools. SLM trained on an $8 ESP32-S3 microcontroller. Runs inference directly on the chip. No cloud dependency. Custom quantized model architecture. Real-time processing at 240MHz clock speed. Demonstrates AI at the absolute edge of computing.

open-edit#

Video editing with AI coding agents requires exporting, uploading, and re-importing. There’s no direct integration between code editors and video manipulation. OpenEdit lets your coding agent edit videos directly. TypeScript-based video editing library. FFmpeg wrapper with programmatic API. Supports cuts, transitions, overlays, and audio mixing. Agent-friendly interface for automated video production.

siematic#

SIEM (Security Information and Event Management) tools are expensive and complex, making them inaccessible to small teams and individual developers. Fair-sourced observability and security platform. Python-based SIEM alternative. Log aggregation, alerting, and threat detection. Self-hosted deployment. Integrates with common log sources. Designed for transparency and community contribution.

listnr#

Meeting transcription tools mix microphone and speaker audio, creating unintelligible transcripts when both sides of a conversation are recorded. macOS meeting transcription that separates mic and speaker audio. Uses Core Audio for device isolation. Local speech-to-text with Whisper. No cloud processing. Real-time transcription with speaker diarization. Swift native implementation.

mcpscore#

MCP (Model Context Protocol) servers need compliance testing to ensure they work correctly with different clients, but there’s no standardized validation tool. Validate any MCP server compliance with the latest MCP version. Automated test suite covering all MCP endpoints. Checks tool definitions, resource access, and prompt handling. CI/CD integration for MCP server development. Python-based with CLI and library modes.

pawmc#

Most LLM assistants require cloud APIs, sending private conversations to third-party servers. Users want a local assistant that learns from their usage patterns. Self-hosted LLM assistant with user-steered compounding context. Builds knowledge graph from interactions. Remembers user preferences and patterns. Local inference with no cloud dependency. Web-based interface. Progressive context accumulation.

pleasantries#

Network handshake and greeting protocols in testing and simulation often require custom implementations that are fragile and hard to maintain. Network protocol greeting and handshake utility. Python library for testing network services. Supports HTTP, TCP, and custom protocols. Automated greeting sequences. Useful for network simulation and testing environments.

adapt#

Creating REST APIs from existing data files requires writing boilerplate code, setting up routing, and building web UIs manually. Automatically turns files into REST APIs, Web UI, and MCP servers. Drop in JSON, CSV, or SQLite files and get a running API. Auto-generates endpoints, documentation, and interactive UI. MCP-compatible for AI agent integration.

paranoia-local#

Code review tools focus on style and formatting but miss security vulnerabilities and logical flaws that automated linters can’t detect. Get a cold, adversarial review of your code. AI-powered code review focused on security and logic flaws. Uses LLM to analyze code from an attacker’s perspective. Identifies injection points, race conditions, and privilege escalation risks. Local processing, no code sent to cloud.

react-mediadrop#

File upload components in React are either too simple (just a file input) or too complex (full-featured uploaders with unnecessary dependencies). Hooks-first headless uploader with pluggable transport. React hooks for file upload with drag-and-drop, progress tracking, and chunk uploads. Transport-agnostic: works with S3, GCS, or custom backends. TypeScript-first with full type safety.

compass#

Understanding large codebases requires expensive cloud-based tools or slow local indexing. AI agents need fast, local code navigation without API costs. Local-first code graph built in Rust for humans and AI agents. Parses code into a navigable graph. Supports 50+ languages. Sub-second queries on million-line codebases. No cloud dependency. MCP server for agent integration. Memory-efficient with incremental updates.

clai#

Using AI in the command line requires complex setups with API keys, custom shells, or heavyweight tools that slow down the terminal workflow. AI for the command line: stdin to LLM to stdout. Pipe any command output through an LLM for analysis. Go binary, single static executable. Works with OpenAI, Anthropic, and local models. Unix philosophy: compose with other tools.

mcpvessel#

Running untrusted MCP servers gives them full access to your system. There’s no sandboxing layer to prevent malicious MCP tools from accessing files or making network calls. Cage untrusted MCP servers in containers, compose them into agents, and share via OCI registries. Signed, sandboxed, no Docker required. Uses containerd for isolation. Egress denied by default. 841 commits. Apache 2.0 licensed.

capshelf#

Teams using AI coding agents accumulate the same skills, settings, and MCP configs across projects. Keeping them in sync manually or via symlinks is fragile. Share agent skills, configs, and MCPs across repos with per-project lockfiles. Git-backed CLI for coding agent configuration. Supports Claude Code, Codex, and Pi extensions. Per-project pinning prevents drift. MIT licensed.

hand-wave#

Sign language translation requires expensive specialized cameras or dedicated hardware. Smart glasses could enable real-time translation but lack accessible toolchains. Detect sign language in real-time with Meta AI glasses and a neural net. Web and iOS apps backed by Python inference. Uses MiCT-RANet for ASL fingerspelling recognition. On-device inference support. 285 commits. MIT licensed.

yanes#

NES and retro console audio synthesis requires complex multi-chip emulation that’s typically locked behind proprietary trackers or requires deep hardware knowledge. Yet Another NES Audio Plugin. Native Linux CLAP instrument for NES and retro console sounds. Polyphonic multi-chip synthesizer covering NES, Game Boy, Genesis, and 15+ expansion chips. Clean-room implementation. Bitwig Studio compatible.

alphai-tui#

Stock market dashboards are either web-based (slow, privacy-leaking) or terminal-based (missing AI analysis and insider trade data). Terminal stock dashboard with live quotes, charts, AI-scored news, and SEC Form 4 insider activity. Rust + ratatui. Multiple data sources including AlphaAI. Real-time candlestick charts with technical indicators. MIT licensed.

Watch#

• Video: https://www.youtube.com/watch?v=mlLT5PUIPXQ
• Cover: https://i.ytimg.com/vi/mlLT5PUIPXQ/hqdefault.jpg

Notes#

Transcript and notes will be added from Forkcast output artifacts.